gpu-kernel-baseline
Pass
Audited by Gen Agent Trust Hub on Sep 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from user-provided PyTorch logic and the local gpu-wiki knowledge base.
- Ingestion points: User-provided logic, kernel demos, and content from the gpu-wiki directory (SKILL.md).
- Boundary markers: No explicit delimiters or instructions are provided to the agent to treat this external data as untrusted or to ignore embedded instructions.
- Capability inventory: The agent has the ability to write files (kernel.py, test_kernel.py), execute shell commands, and run generated code.
- Sanitization: There is no requirement or mechanism for sanitizing the input data before processing or code generation.
- [COMMAND_EXECUTION]: The skill executes multiple local scripts via the command line, including tools for querying the wiki, calculating utilization, and managing iteration memory (Phase 2 and Phase 4).
- [DYNAMIC_EXECUTION]: The skill generates Python code (kernel.py and test_kernel.py) and executes it within a mandatory sandbox (tools/sandbox.py). This represents script generation and execution at runtime.
Audit Metadata